[Paper Review] Simulation of Open Quantum Dynamics with Bootstrap-Based Long Short-Term Memory Recurrent Neural Network

Kunni Lin, Jiawei Peng+2 more|arXiv (Cornell University)|Aug 3, 2021

Spectroscopy and Quantum Chemical Studies67 references50 citations

TL;DR

This paper proposes a bootstrap-based long short-term memory recurrent neural network (LSTM-NN) to simulate long-time open quantum dynamics with high accuracy and low computational cost. By training on early-stage quantum evolution data from the multilayer multiconfigurational time-dependent Hartree (ML-MCTDH) method, the LSTM-NN ensemble enables reliable long-time propagation while the bootstrap method quantifies prediction uncertainty, achieving strong consistency with numerically exact results.

ABSTRACT

The recurrent neural network with the long short-term memory cell (LSTM-NN) is employed to simulate the long-time dynamics of open quantum system. The bootstrap method is applied in the LSTM-NN construction and prediction, which provides a Monte-Carlo estimation of forecasting confidence interval. Within this approach, a large number of LSTM-NNs are constructed by resampling time-series sequences that were obtained from the early-stage quantum evolution given by numerically-exact multilayer multiconfigurational time-dependent Hartree method. The built LSTM-NN ensemble is used for the reliable propagation of the long-time quantum dynamics and the simulated result is highly consistent with the exact evolution. The forecasting uncertainty that partially reflects the reliability of the LSTM-NN prediction is also given. This demonstrates the bootstrap-based LSTM-NN approach is a practical and powerful tool to propagate the long-time quantum dynamics of open systems with high accuracy and low computational cost.

Motivation & Objective

• To develop a computationally efficient and accurate method for simulating long-time open quantum dynamics beyond the reach of conventional approaches.
• To address the lack of reliable uncertainty estimation in machine learning-based quantum dynamics predictions.
• To combine the expressive power of LSTM-NNs with the statistical robustness of bootstrap resampling for enhanced forecasting reliability.
• To demonstrate the method's effectiveness on non-Markovian, strong-coupling, and low-temperature spin-boson models.

Proposed method

• A large ensemble of LSTM-NNs is trained on short-time quantum evolution data obtained via the numerically exact ML-MCTDH method.
• Bootstrap resampling is applied to the training data to generate multiple independent LSTM-NN models, enabling statistical estimation of prediction confidence intervals.
• The input to each LSTM-NN consists of time-series sequences of reduced density matrix elements over a defined window length L, with the output predicting the next time point.
• Model hyperparameters (number of layers, neurons, sequence length L) are optimized via grid search to minimize validation error.
• The final prediction is obtained as the ensemble average of all bootstrap-trained LSTM-NNs, with uncertainty quantified by the distribution of predictions.
• For challenging cases (e.g., low temperature, strong coupling), additional model selection is applied to retain only low-validation-error networks before bootstrap.

Experimental results

Research questions

• RQ1 Can an LSTM-NN ensemble trained on short-time quantum evolution accurately predict long-time open quantum dynamics?
• RQ2 Can the bootstrap method reliably quantify the uncertainty of LSTM-NN predictions in quantum dynamics simulations?
• RQ3 How does the performance of the bootstrap-LSTM-NN method compare to numerically exact methods like hybrid stochastic-deterministic HEOM?
• RQ4 What impact do missing dynamical features (e.g., off-diagonal density matrix elements) have on prediction reliability and uncertainty?
• RQ5 Can the method handle strongly non-Markovian dynamics under low-temperature and strong system-bath coupling conditions?

Key findings

• The bootstrap-LSTM-NN method achieves high consistency with numerically exact ML-MCTDH and hybrid stochastic-deterministic HEOM results across various system parameters.
• The prediction uncertainty, derived from the bootstrap ensemble, increases over time and correlates with regions of high dynamical complexity, such as non-Markovian oscillations.
• For the spin-boson model at 10 K and strong coupling, the ensemble average prediction closely matches the long-time limit obtained from imaginary-time path integral simulations.
• The method successfully captures non-Markovian dynamics, including population oscillations and coherence effects, even in challenging low-temperature regimes.
• The visible increase in prediction uncertainty at 10 K is attributed to the omission of off-diagonal density matrix elements in the model input, highlighting a key limitation.
• The approach significantly reduces computational cost compared to exact methods while maintaining high accuracy and providing intrinsic uncertainty quantification.